blob: 883f7208901585ab1ee02891a3ad5b67599a4243 [file] [log] [blame]
Jens Axboe320ae512013-10-24 09:20:05 +01001#include <linux/kernel.h>
2#include <linux/module.h>
3#include <linux/backing-dev.h>
4#include <linux/bio.h>
5#include <linux/blkdev.h>
6#include <linux/mm.h>
7#include <linux/init.h>
8#include <linux/slab.h>
9#include <linux/workqueue.h>
10#include <linux/smp.h>
11#include <linux/llist.h>
12#include <linux/list_sort.h>
13#include <linux/cpu.h>
14#include <linux/cache.h>
15#include <linux/sched/sysctl.h>
16#include <linux/delay.h>
17
18#include <trace/events/block.h>
19
20#include <linux/blk-mq.h>
21#include "blk.h"
22#include "blk-mq.h"
23#include "blk-mq-tag.h"
24
25static DEFINE_MUTEX(all_q_mutex);
26static LIST_HEAD(all_q_list);
27
28static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
29
Jens Axboe320ae512013-10-24 09:20:05 +010030static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
31 unsigned int cpu)
32{
33 return per_cpu_ptr(q->queue_ctx, cpu);
34}
35
36/*
37 * This assumes per-cpu software queueing queues. They could be per-node
38 * as well, for instance. For now this is hardcoded as-is. Note that we don't
39 * care about preemption, since we know the ctx's are persistent. This does
40 * mean that we can't rely on ctx always matching the currently running CPU.
41 */
42static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
43{
44 return __blk_mq_get_ctx(q, get_cpu());
45}
46
47static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
48{
49 put_cpu();
50}
51
52/*
53 * Check if any of the ctx's have pending work in this hardware queue
54 */
55static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
56{
57 unsigned int i;
58
59 for (i = 0; i < hctx->nr_ctx_map; i++)
60 if (hctx->ctx_map[i])
61 return true;
62
63 return false;
64}
65
66/*
67 * Mark this ctx as having pending work in this hardware queue
68 */
69static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
70 struct blk_mq_ctx *ctx)
71{
72 if (!test_bit(ctx->index_hw, hctx->ctx_map))
73 set_bit(ctx->index_hw, hctx->ctx_map);
74}
75
Christoph Hellwigfd694132014-02-20 15:32:36 -080076static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
77 gfp_t gfp, bool reserved)
Jens Axboe320ae512013-10-24 09:20:05 +010078{
79 struct request *rq;
80 unsigned int tag;
81
82 tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
83 if (tag != BLK_MQ_TAG_FAIL) {
84 rq = hctx->rqs[tag];
85 rq->tag = tag;
86
87 return rq;
88 }
89
90 return NULL;
91}
92
93static int blk_mq_queue_enter(struct request_queue *q)
94{
95 int ret;
96
97 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
98 smp_wmb();
99 /* we have problems to freeze the queue if it's initializing */
100 if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
101 return 0;
102
103 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
104
105 spin_lock_irq(q->queue_lock);
106 ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
Ming Lei43a5e4e2013-12-26 21:31:35 +0800107 !blk_queue_bypass(q) || blk_queue_dying(q),
108 *q->queue_lock);
Jens Axboe320ae512013-10-24 09:20:05 +0100109 /* inc usage with lock hold to avoid freeze_queue runs here */
Ming Lei43a5e4e2013-12-26 21:31:35 +0800110 if (!ret && !blk_queue_dying(q))
Jens Axboe320ae512013-10-24 09:20:05 +0100111 __percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
Ming Lei43a5e4e2013-12-26 21:31:35 +0800112 else if (blk_queue_dying(q))
113 ret = -ENODEV;
Jens Axboe320ae512013-10-24 09:20:05 +0100114 spin_unlock_irq(q->queue_lock);
115
116 return ret;
117}
118
119static void blk_mq_queue_exit(struct request_queue *q)
120{
121 __percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
122}
123
Ming Lei43a5e4e2013-12-26 21:31:35 +0800124static void __blk_mq_drain_queue(struct request_queue *q)
125{
126 while (true) {
127 s64 count;
128
129 spin_lock_irq(q->queue_lock);
130 count = percpu_counter_sum(&q->mq_usage_counter);
131 spin_unlock_irq(q->queue_lock);
132
133 if (count == 0)
134 break;
135 blk_mq_run_queues(q, false);
136 msleep(10);
137 }
138}
139
Jens Axboe320ae512013-10-24 09:20:05 +0100140/*
141 * Guarantee no request is in use, so we can change any data structure of
142 * the queue afterward.
143 */
144static void blk_mq_freeze_queue(struct request_queue *q)
145{
146 bool drain;
147
148 spin_lock_irq(q->queue_lock);
149 drain = !q->bypass_depth++;
150 queue_flag_set(QUEUE_FLAG_BYPASS, q);
151 spin_unlock_irq(q->queue_lock);
152
Ming Lei43a5e4e2013-12-26 21:31:35 +0800153 if (drain)
154 __blk_mq_drain_queue(q);
155}
Jens Axboe320ae512013-10-24 09:20:05 +0100156
Ming Lei43a5e4e2013-12-26 21:31:35 +0800157void blk_mq_drain_queue(struct request_queue *q)
158{
159 __blk_mq_drain_queue(q);
Jens Axboe320ae512013-10-24 09:20:05 +0100160}
161
162static void blk_mq_unfreeze_queue(struct request_queue *q)
163{
164 bool wake = false;
165
166 spin_lock_irq(q->queue_lock);
167 if (!--q->bypass_depth) {
168 queue_flag_clear(QUEUE_FLAG_BYPASS, q);
169 wake = true;
170 }
171 WARN_ON_ONCE(q->bypass_depth < 0);
172 spin_unlock_irq(q->queue_lock);
173 if (wake)
174 wake_up_all(&q->mq_freeze_wq);
175}
176
177bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
178{
179 return blk_mq_has_free_tags(hctx->tags);
180}
181EXPORT_SYMBOL(blk_mq_can_queue);
182
Jens Axboe94eddfb2013-11-19 09:25:07 -0700183static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
184 struct request *rq, unsigned int rw_flags)
Jens Axboe320ae512013-10-24 09:20:05 +0100185{
Jens Axboe94eddfb2013-11-19 09:25:07 -0700186 if (blk_queue_io_stat(q))
187 rw_flags |= REQ_IO_STAT;
188
Jens Axboe320ae512013-10-24 09:20:05 +0100189 rq->mq_ctx = ctx;
190 rq->cmd_flags = rw_flags;
Ming Lei0fec08b2014-01-03 10:00:08 -0700191 rq->start_time = jiffies;
192 set_start_time_ns(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
194}
195
Jens Axboe320ae512013-10-24 09:20:05 +0100196static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
197 int rw, gfp_t gfp,
198 bool reserved)
199{
200 struct request *rq;
201
202 do {
203 struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
204 struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
205
Christoph Hellwig18741982014-02-10 09:29:00 -0700206 rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
Jens Axboe320ae512013-10-24 09:20:05 +0100207 if (rq) {
Jens Axboe94eddfb2013-11-19 09:25:07 -0700208 blk_mq_rq_ctx_init(q, ctx, rq, rw);
Jens Axboe320ae512013-10-24 09:20:05 +0100209 break;
Jeff Moyer959a35f2013-12-03 14:23:00 -0700210 }
Jens Axboe320ae512013-10-24 09:20:05 +0100211
212 blk_mq_put_ctx(ctx);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700213 if (!(gfp & __GFP_WAIT))
214 break;
215
Jens Axboe320ae512013-10-24 09:20:05 +0100216 __blk_mq_run_hw_queue(hctx);
217 blk_mq_wait_for_tags(hctx->tags);
218 } while (1);
219
220 return rq;
221}
222
Christoph Hellwig18741982014-02-10 09:29:00 -0700223struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
Jens Axboe320ae512013-10-24 09:20:05 +0100224{
225 struct request *rq;
226
227 if (blk_mq_queue_enter(q))
228 return NULL;
229
Christoph Hellwig18741982014-02-10 09:29:00 -0700230 rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700231 if (rq)
232 blk_mq_put_ctx(rq->mq_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100233 return rq;
234}
235
236struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
237 gfp_t gfp)
238{
239 struct request *rq;
240
241 if (blk_mq_queue_enter(q))
242 return NULL;
243
244 rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
Jeff Moyer959a35f2013-12-03 14:23:00 -0700245 if (rq)
246 blk_mq_put_ctx(rq->mq_ctx);
Jens Axboe320ae512013-10-24 09:20:05 +0100247 return rq;
248}
249EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
250
251/*
252 * Re-init and set pdu, if we have it
253 */
Christoph Hellwig18741982014-02-10 09:29:00 -0700254void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100255{
256 blk_rq_init(hctx->queue, rq);
257
258 if (hctx->cmd_size)
259 rq->special = blk_mq_rq_to_pdu(rq);
260}
261
262static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
263 struct blk_mq_ctx *ctx, struct request *rq)
264{
265 const int tag = rq->tag;
266 struct request_queue *q = rq->q;
267
268 blk_mq_rq_init(hctx, rq);
269 blk_mq_put_tag(hctx->tags, tag);
270
271 blk_mq_queue_exit(q);
272}
273
274void blk_mq_free_request(struct request *rq)
275{
276 struct blk_mq_ctx *ctx = rq->mq_ctx;
277 struct blk_mq_hw_ctx *hctx;
278 struct request_queue *q = rq->q;
279
280 ctx->rq_completed[rq_is_sync(rq)]++;
281
282 hctx = q->mq_ops->map_queue(q, ctx->cpu);
283 __blk_mq_free_request(hctx, ctx, rq);
284}
285
Christoph Hellwigd6a25b32014-02-20 15:32:38 -0800286bool blk_mq_end_io_partial(struct request *rq, int error, unsigned int nr_bytes)
Jens Axboe320ae512013-10-24 09:20:05 +0100287{
Christoph Hellwigd6a25b32014-02-20 15:32:38 -0800288 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
289 return true;
Jens Axboe320ae512013-10-24 09:20:05 +0100290
Ming Lei0d11e6a2013-12-05 10:50:39 -0700291 blk_account_io_done(rq);
292
Jens Axboe320ae512013-10-24 09:20:05 +0100293 if (rq->end_io)
294 rq->end_io(rq, error);
295 else
296 blk_mq_free_request(rq);
Christoph Hellwigd6a25b32014-02-20 15:32:38 -0800297 return false;
Jens Axboe320ae512013-10-24 09:20:05 +0100298}
Christoph Hellwigd6a25b32014-02-20 15:32:38 -0800299EXPORT_SYMBOL(blk_mq_end_io_partial);
Jens Axboe320ae512013-10-24 09:20:05 +0100300
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800301static void __blk_mq_complete_request_remote(void *data)
Jens Axboe320ae512013-10-24 09:20:05 +0100302{
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800303 struct request *rq = data;
Jens Axboe320ae512013-10-24 09:20:05 +0100304
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800305 rq->q->softirq_done_fn(rq);
Jens Axboe320ae512013-10-24 09:20:05 +0100306}
307
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800308void __blk_mq_complete_request(struct request *rq)
Jens Axboe320ae512013-10-24 09:20:05 +0100309{
310 struct blk_mq_ctx *ctx = rq->mq_ctx;
311 int cpu;
312
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800313 if (!ctx->ipi_redirect) {
314 rq->q->softirq_done_fn(rq);
315 return;
316 }
Jens Axboe320ae512013-10-24 09:20:05 +0100317
318 cpu = get_cpu();
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800319 if (cpu != ctx->cpu && cpu_online(ctx->cpu)) {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800320 rq->csd.func = __blk_mq_complete_request_remote;
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800321 rq->csd.info = rq;
322 rq->csd.flags = 0;
323 __smp_call_function_single(ctx->cpu, &rq->csd, 0);
324 } else {
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800325 rq->q->softirq_done_fn(rq);
Christoph Hellwig3d6efbf2014-01-08 09:33:37 -0800326 }
Jens Axboe320ae512013-10-24 09:20:05 +0100327 put_cpu();
328}
Christoph Hellwig30a91cb2014-02-10 03:24:38 -0800329
330/**
331 * blk_mq_complete_request - end I/O on a request
332 * @rq: the request being processed
333 *
334 * Description:
335 * Ends all I/O on a request. It does not handle partial completions.
336 * The actual completion happens out-of-order, through a IPI handler.
337 **/
338void blk_mq_complete_request(struct request *rq)
339{
340 if (unlikely(blk_should_fake_timeout(rq->q)))
341 return;
342 if (!blk_mark_rq_complete(rq))
343 __blk_mq_complete_request(rq);
344}
345EXPORT_SYMBOL(blk_mq_complete_request);
Jens Axboe320ae512013-10-24 09:20:05 +0100346
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800347static void blk_mq_start_request(struct request *rq, bool last)
Jens Axboe320ae512013-10-24 09:20:05 +0100348{
349 struct request_queue *q = rq->q;
350
351 trace_block_rq_issue(q, rq);
352
353 /*
354 * Just mark start time and set the started bit. Due to memory
355 * ordering, we know we'll see the correct deadline as long as
356 * REQ_ATOMIC_STARTED is seen.
357 */
358 rq->deadline = jiffies + q->rq_timeout;
359 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800360
361 if (q->dma_drain_size && blk_rq_bytes(rq)) {
362 /*
363 * Make sure space for the drain appears. We know we can do
364 * this because max_hw_segments has been adjusted to be one
365 * fewer than the device can handle.
366 */
367 rq->nr_phys_segments++;
368 }
369
370 /*
371 * Flag the last request in the series so that drivers know when IO
372 * should be kicked off, if they don't do it on a per-request basis.
373 *
374 * Note: the flag isn't the only condition drivers should do kick off.
375 * If drive is busy, the last request might not have the bit set.
376 */
377 if (last)
378 rq->cmd_flags |= REQ_END;
Jens Axboe320ae512013-10-24 09:20:05 +0100379}
380
381static void blk_mq_requeue_request(struct request *rq)
382{
383 struct request_queue *q = rq->q;
384
385 trace_block_rq_requeue(q, rq);
386 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800387
388 rq->cmd_flags &= ~REQ_END;
389
390 if (q->dma_drain_size && blk_rq_bytes(rq))
391 rq->nr_phys_segments--;
Jens Axboe320ae512013-10-24 09:20:05 +0100392}
393
394struct blk_mq_timeout_data {
395 struct blk_mq_hw_ctx *hctx;
396 unsigned long *next;
397 unsigned int *next_set;
398};
399
400static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
401{
402 struct blk_mq_timeout_data *data = __data;
403 struct blk_mq_hw_ctx *hctx = data->hctx;
404 unsigned int tag;
405
406 /* It may not be in flight yet (this is where
407 * the REQ_ATOMIC_STARTED flag comes in). The requests are
408 * statically allocated, so we know it's always safe to access the
409 * memory associated with a bit offset into ->rqs[].
410 */
411 tag = 0;
412 do {
413 struct request *rq;
414
415 tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
416 if (tag >= hctx->queue_depth)
417 break;
418
419 rq = hctx->rqs[tag++];
420
421 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
422 continue;
423
424 blk_rq_check_expired(rq, data->next, data->next_set);
425 } while (1);
426}
427
428static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
429 unsigned long *next,
430 unsigned int *next_set)
431{
432 struct blk_mq_timeout_data data = {
433 .hctx = hctx,
434 .next = next,
435 .next_set = next_set,
436 };
437
438 /*
439 * Ask the tagging code to iterate busy requests, so we can
440 * check them for timeout.
441 */
442 blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
443}
444
445static void blk_mq_rq_timer(unsigned long data)
446{
447 struct request_queue *q = (struct request_queue *) data;
448 struct blk_mq_hw_ctx *hctx;
449 unsigned long next = 0;
450 int i, next_set = 0;
451
452 queue_for_each_hw_ctx(q, hctx, i)
453 blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
454
455 if (next_set)
456 mod_timer(&q->timeout, round_jiffies_up(next));
457}
458
459/*
460 * Reverse check our software queue for entries that we could potentially
461 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
462 * too much time checking for merges.
463 */
464static bool blk_mq_attempt_merge(struct request_queue *q,
465 struct blk_mq_ctx *ctx, struct bio *bio)
466{
467 struct request *rq;
468 int checked = 8;
469
470 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
471 int el_ret;
472
473 if (!checked--)
474 break;
475
476 if (!blk_rq_merge_ok(rq, bio))
477 continue;
478
479 el_ret = blk_try_merge(rq, bio);
480 if (el_ret == ELEVATOR_BACK_MERGE) {
481 if (bio_attempt_back_merge(q, rq, bio)) {
482 ctx->rq_merged++;
483 return true;
484 }
485 break;
486 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
487 if (bio_attempt_front_merge(q, rq, bio)) {
488 ctx->rq_merged++;
489 return true;
490 }
491 break;
492 }
493 }
494
495 return false;
496}
497
498void blk_mq_add_timer(struct request *rq)
499{
500 __blk_add_timer(rq, NULL);
501}
502
503/*
504 * Run this hardware queue, pulling any software queues mapped to it in.
505 * Note that this function currently has various problems around ordering
506 * of IO. In particular, we'd like FIFO behaviour on handling existing
507 * items on the hctx->dispatch list. Ignore that for now.
508 */
509static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
510{
511 struct request_queue *q = hctx->queue;
512 struct blk_mq_ctx *ctx;
513 struct request *rq;
514 LIST_HEAD(rq_list);
515 int bit, queued;
516
517 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
518 return;
519
520 hctx->run++;
521
522 /*
523 * Touch any software queue that has pending entries.
524 */
525 for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
526 clear_bit(bit, hctx->ctx_map);
527 ctx = hctx->ctxs[bit];
528 BUG_ON(bit != ctx->index_hw);
529
530 spin_lock(&ctx->lock);
531 list_splice_tail_init(&ctx->rq_list, &rq_list);
532 spin_unlock(&ctx->lock);
533 }
534
535 /*
536 * If we have previous entries on our dispatch list, grab them
537 * and stuff them at the front for more fair dispatch.
538 */
539 if (!list_empty_careful(&hctx->dispatch)) {
540 spin_lock(&hctx->lock);
541 if (!list_empty(&hctx->dispatch))
542 list_splice_init(&hctx->dispatch, &rq_list);
543 spin_unlock(&hctx->lock);
544 }
545
546 /*
547 * Delete and return all entries from our dispatch list
548 */
549 queued = 0;
550
551 /*
552 * Now process all the entries, sending them to the driver.
553 */
554 while (!list_empty(&rq_list)) {
555 int ret;
556
557 rq = list_first_entry(&rq_list, struct request, queuelist);
558 list_del_init(&rq->queuelist);
Jens Axboe320ae512013-10-24 09:20:05 +0100559
Christoph Hellwig49f5baa2014-02-11 08:27:14 -0800560 blk_mq_start_request(rq, list_empty(&rq_list));
Jens Axboe320ae512013-10-24 09:20:05 +0100561
562 ret = q->mq_ops->queue_rq(hctx, rq);
563 switch (ret) {
564 case BLK_MQ_RQ_QUEUE_OK:
565 queued++;
566 continue;
567 case BLK_MQ_RQ_QUEUE_BUSY:
568 /*
569 * FIXME: we should have a mechanism to stop the queue
570 * like blk_stop_queue, otherwise we will waste cpu
571 * time
572 */
573 list_add(&rq->queuelist, &rq_list);
574 blk_mq_requeue_request(rq);
575 break;
576 default:
577 pr_err("blk-mq: bad return on queue: %d\n", ret);
Jens Axboe320ae512013-10-24 09:20:05 +0100578 case BLK_MQ_RQ_QUEUE_ERROR:
Christoph Hellwig1e93b8c2014-02-11 08:27:13 -0800579 rq->errors = -EIO;
Jens Axboe320ae512013-10-24 09:20:05 +0100580 blk_mq_end_io(rq, rq->errors);
581 break;
582 }
583
584 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
585 break;
586 }
587
588 if (!queued)
589 hctx->dispatched[0]++;
590 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
591 hctx->dispatched[ilog2(queued) + 1]++;
592
593 /*
594 * Any items that need requeuing? Stuff them into hctx->dispatch,
595 * that is where we will continue on next queue run.
596 */
597 if (!list_empty(&rq_list)) {
598 spin_lock(&hctx->lock);
599 list_splice(&rq_list, &hctx->dispatch);
600 spin_unlock(&hctx->lock);
601 }
602}
603
604void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
605{
606 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
607 return;
608
609 if (!async)
610 __blk_mq_run_hw_queue(hctx);
611 else {
612 struct request_queue *q = hctx->queue;
613
614 kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
615 }
616}
617
618void blk_mq_run_queues(struct request_queue *q, bool async)
619{
620 struct blk_mq_hw_ctx *hctx;
621 int i;
622
623 queue_for_each_hw_ctx(q, hctx, i) {
624 if ((!blk_mq_hctx_has_pending(hctx) &&
625 list_empty_careful(&hctx->dispatch)) ||
626 test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
627 continue;
628
629 blk_mq_run_hw_queue(hctx, async);
630 }
631}
632EXPORT_SYMBOL(blk_mq_run_queues);
633
634void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
635{
636 cancel_delayed_work(&hctx->delayed_work);
637 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
638}
639EXPORT_SYMBOL(blk_mq_stop_hw_queue);
640
Christoph Hellwig280d45f2013-10-25 14:45:58 +0100641void blk_mq_stop_hw_queues(struct request_queue *q)
642{
643 struct blk_mq_hw_ctx *hctx;
644 int i;
645
646 queue_for_each_hw_ctx(q, hctx, i)
647 blk_mq_stop_hw_queue(hctx);
648}
649EXPORT_SYMBOL(blk_mq_stop_hw_queues);
650
Jens Axboe320ae512013-10-24 09:20:05 +0100651void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
652{
653 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
654 __blk_mq_run_hw_queue(hctx);
655}
656EXPORT_SYMBOL(blk_mq_start_hw_queue);
657
658void blk_mq_start_stopped_hw_queues(struct request_queue *q)
659{
660 struct blk_mq_hw_ctx *hctx;
661 int i;
662
663 queue_for_each_hw_ctx(q, hctx, i) {
664 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
665 continue;
666
667 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
668 blk_mq_run_hw_queue(hctx, true);
669 }
670}
671EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
672
673static void blk_mq_work_fn(struct work_struct *work)
674{
675 struct blk_mq_hw_ctx *hctx;
676
677 hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
678 __blk_mq_run_hw_queue(hctx);
679}
680
681static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800682 struct request *rq, bool at_head)
Jens Axboe320ae512013-10-24 09:20:05 +0100683{
684 struct blk_mq_ctx *ctx = rq->mq_ctx;
685
Jens Axboe01b983c2013-11-19 18:59:10 -0700686 trace_block_rq_insert(hctx->queue, rq);
687
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800688 if (at_head)
689 list_add(&rq->queuelist, &ctx->rq_list);
690 else
691 list_add_tail(&rq->queuelist, &ctx->rq_list);
Jens Axboe320ae512013-10-24 09:20:05 +0100692 blk_mq_hctx_mark_pending(hctx, ctx);
693
694 /*
695 * We do this early, to ensure we are on the right CPU.
696 */
697 blk_mq_add_timer(rq);
698}
699
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800700void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
701 bool async)
Jens Axboe320ae512013-10-24 09:20:05 +0100702{
703 struct request_queue *q = rq->q;
704 struct blk_mq_hw_ctx *hctx;
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800705 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100706
707 current_ctx = blk_mq_get_ctx(q);
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800708 if (!cpu_online(ctx->cpu))
709 rq->mq_ctx = ctx = current_ctx;
Jens Axboe320ae512013-10-24 09:20:05 +0100710
Jens Axboe320ae512013-10-24 09:20:05 +0100711 hctx = q->mq_ops->map_queue(q, ctx->cpu);
712
Christoph Hellwigfeb71da2014-02-20 15:32:37 -0800713 if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA) &&
714 !(rq->cmd_flags & (REQ_FLUSH_SEQ))) {
715 blk_insert_flush(rq);
716 } else {
717 spin_lock(&ctx->lock);
718 __blk_mq_insert_request(hctx, rq, at_head);
719 spin_unlock(&ctx->lock);
720 }
Jens Axboe320ae512013-10-24 09:20:05 +0100721
722 blk_mq_put_ctx(current_ctx);
723
724 if (run_queue)
725 blk_mq_run_hw_queue(hctx, async);
726}
727
728static void blk_mq_insert_requests(struct request_queue *q,
729 struct blk_mq_ctx *ctx,
730 struct list_head *list,
731 int depth,
732 bool from_schedule)
733
734{
735 struct blk_mq_hw_ctx *hctx;
736 struct blk_mq_ctx *current_ctx;
737
738 trace_block_unplug(q, depth, !from_schedule);
739
740 current_ctx = blk_mq_get_ctx(q);
741
742 if (!cpu_online(ctx->cpu))
743 ctx = current_ctx;
744 hctx = q->mq_ops->map_queue(q, ctx->cpu);
745
746 /*
747 * preemption doesn't flush plug list, so it's possible ctx->cpu is
748 * offline now
749 */
750 spin_lock(&ctx->lock);
751 while (!list_empty(list)) {
752 struct request *rq;
753
754 rq = list_first_entry(list, struct request, queuelist);
755 list_del_init(&rq->queuelist);
756 rq->mq_ctx = ctx;
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800757 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100758 }
759 spin_unlock(&ctx->lock);
760
761 blk_mq_put_ctx(current_ctx);
762
763 blk_mq_run_hw_queue(hctx, from_schedule);
764}
765
766static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
767{
768 struct request *rqa = container_of(a, struct request, queuelist);
769 struct request *rqb = container_of(b, struct request, queuelist);
770
771 return !(rqa->mq_ctx < rqb->mq_ctx ||
772 (rqa->mq_ctx == rqb->mq_ctx &&
773 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
774}
775
776void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
777{
778 struct blk_mq_ctx *this_ctx;
779 struct request_queue *this_q;
780 struct request *rq;
781 LIST_HEAD(list);
782 LIST_HEAD(ctx_list);
783 unsigned int depth;
784
785 list_splice_init(&plug->mq_list, &list);
786
787 list_sort(NULL, &list, plug_ctx_cmp);
788
789 this_q = NULL;
790 this_ctx = NULL;
791 depth = 0;
792
793 while (!list_empty(&list)) {
794 rq = list_entry_rq(list.next);
795 list_del_init(&rq->queuelist);
796 BUG_ON(!rq->q);
797 if (rq->mq_ctx != this_ctx) {
798 if (this_ctx) {
799 blk_mq_insert_requests(this_q, this_ctx,
800 &ctx_list, depth,
801 from_schedule);
802 }
803
804 this_ctx = rq->mq_ctx;
805 this_q = rq->q;
806 depth = 0;
807 }
808
809 depth++;
810 list_add_tail(&rq->queuelist, &ctx_list);
811 }
812
813 /*
814 * If 'this_ctx' is set, we know we have entries to complete
815 * on 'ctx_list'. Do those.
816 */
817 if (this_ctx) {
818 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
819 from_schedule);
820 }
821}
822
823static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
824{
825 init_request_from_bio(rq, bio);
826 blk_account_io_start(rq, 1);
827}
828
829static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
830{
831 struct blk_mq_hw_ctx *hctx;
832 struct blk_mq_ctx *ctx;
833 const int is_sync = rw_is_sync(bio->bi_rw);
834 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
835 int rw = bio_data_dir(bio);
836 struct request *rq;
837 unsigned int use_plug, request_count = 0;
838
839 /*
840 * If we have multiple hardware queues, just go directly to
841 * one of those for sync IO.
842 */
843 use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
844
845 blk_queue_bounce(q, &bio);
846
Nicholas Bellinger14ec77f2014-02-07 13:45:39 -0700847 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
848 bio_endio(bio, -EIO);
849 return;
850 }
851
Jens Axboe320ae512013-10-24 09:20:05 +0100852 if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
853 return;
854
855 if (blk_mq_queue_enter(q)) {
856 bio_endio(bio, -EIO);
857 return;
858 }
859
860 ctx = blk_mq_get_ctx(q);
861 hctx = q->mq_ops->map_queue(q, ctx->cpu);
862
Shaohua Li739c3ee2014-02-19 20:20:21 +0800863 if (is_sync)
864 rw |= REQ_SYNC;
Jens Axboe320ae512013-10-24 09:20:05 +0100865 trace_block_getrq(q, bio, rw);
Christoph Hellwig18741982014-02-10 09:29:00 -0700866 rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100867 if (likely(rq))
Christoph Hellwig18741982014-02-10 09:29:00 -0700868 blk_mq_rq_ctx_init(q, ctx, rq, rw);
Jens Axboe320ae512013-10-24 09:20:05 +0100869 else {
870 blk_mq_put_ctx(ctx);
871 trace_block_sleeprq(q, bio, rw);
Christoph Hellwig18741982014-02-10 09:29:00 -0700872 rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
873 false);
Jens Axboe320ae512013-10-24 09:20:05 +0100874 ctx = rq->mq_ctx;
875 hctx = q->mq_ops->map_queue(q, ctx->cpu);
876 }
877
878 hctx->queued++;
879
880 if (unlikely(is_flush_fua)) {
881 blk_mq_bio_to_request(rq, bio);
882 blk_mq_put_ctx(ctx);
883 blk_insert_flush(rq);
884 goto run_queue;
885 }
886
887 /*
888 * A task plug currently exists. Since this is completely lockless,
889 * utilize that to temporarily store requests until the task is
890 * either done or scheduled away.
891 */
892 if (use_plug) {
893 struct blk_plug *plug = current->plug;
894
895 if (plug) {
896 blk_mq_bio_to_request(rq, bio);
Shaohua Li92f399c2013-10-29 12:01:03 -0600897 if (list_empty(&plug->mq_list))
Jens Axboe320ae512013-10-24 09:20:05 +0100898 trace_block_plug(q);
899 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
900 blk_flush_plug_list(plug, false);
901 trace_block_plug(q);
902 }
903 list_add_tail(&rq->queuelist, &plug->mq_list);
904 blk_mq_put_ctx(ctx);
905 return;
906 }
907 }
908
909 spin_lock(&ctx->lock);
910
911 if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
912 blk_mq_attempt_merge(q, ctx, bio))
913 __blk_mq_free_request(hctx, ctx, rq);
914 else {
915 blk_mq_bio_to_request(rq, bio);
Christoph Hellwig72a0a362014-02-07 10:22:36 -0800916 __blk_mq_insert_request(hctx, rq, false);
Jens Axboe320ae512013-10-24 09:20:05 +0100917 }
918
919 spin_unlock(&ctx->lock);
920 blk_mq_put_ctx(ctx);
921
922 /*
923 * For a SYNC request, send it to the hardware immediately. For an
924 * ASYNC request, just ensure that we run it later on. The latter
925 * allows for merging opportunities and more efficient dispatching.
926 */
927run_queue:
928 blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
929}
930
931/*
932 * Default mapping to a software queue, since we use one per CPU.
933 */
934struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
935{
936 return q->queue_hw_ctx[q->mq_map[cpu]];
937}
938EXPORT_SYMBOL(blk_mq_map_queue);
939
940struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
941 unsigned int hctx_index)
942{
943 return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
944 GFP_KERNEL | __GFP_ZERO, reg->numa_node);
945}
946EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
947
948void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
949 unsigned int hctx_index)
950{
951 kfree(hctx);
952}
953EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
954
955static void blk_mq_hctx_notify(void *data, unsigned long action,
956 unsigned int cpu)
957{
958 struct blk_mq_hw_ctx *hctx = data;
959 struct blk_mq_ctx *ctx;
960 LIST_HEAD(tmp);
961
962 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
963 return;
964
965 /*
966 * Move ctx entries to new CPU, if this one is going away.
967 */
968 ctx = __blk_mq_get_ctx(hctx->queue, cpu);
969
970 spin_lock(&ctx->lock);
971 if (!list_empty(&ctx->rq_list)) {
972 list_splice_init(&ctx->rq_list, &tmp);
973 clear_bit(ctx->index_hw, hctx->ctx_map);
974 }
975 spin_unlock(&ctx->lock);
976
977 if (list_empty(&tmp))
978 return;
979
980 ctx = blk_mq_get_ctx(hctx->queue);
981 spin_lock(&ctx->lock);
982
983 while (!list_empty(&tmp)) {
984 struct request *rq;
985
986 rq = list_first_entry(&tmp, struct request, queuelist);
987 rq->mq_ctx = ctx;
988 list_move_tail(&rq->queuelist, &ctx->rq_list);
989 }
990
991 blk_mq_hctx_mark_pending(hctx, ctx);
992
993 spin_unlock(&ctx->lock);
994 blk_mq_put_ctx(ctx);
995}
996
997static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
998 void (*init)(void *, struct blk_mq_hw_ctx *,
999 struct request *, unsigned int),
1000 void *data)
1001{
1002 unsigned int i;
1003
1004 for (i = 0; i < hctx->queue_depth; i++) {
1005 struct request *rq = hctx->rqs[i];
1006
1007 init(data, hctx, rq, i);
1008 }
1009}
1010
1011void blk_mq_init_commands(struct request_queue *q,
1012 void (*init)(void *, struct blk_mq_hw_ctx *,
1013 struct request *, unsigned int),
1014 void *data)
1015{
1016 struct blk_mq_hw_ctx *hctx;
1017 unsigned int i;
1018
1019 queue_for_each_hw_ctx(q, hctx, i)
1020 blk_mq_init_hw_commands(hctx, init, data);
1021}
1022EXPORT_SYMBOL(blk_mq_init_commands);
1023
1024static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
1025{
1026 struct page *page;
1027
1028 while (!list_empty(&hctx->page_list)) {
Dave Hansen67534712014-01-08 20:17:46 -07001029 page = list_first_entry(&hctx->page_list, struct page, lru);
1030 list_del_init(&page->lru);
Jens Axboe320ae512013-10-24 09:20:05 +01001031 __free_pages(page, page->private);
1032 }
1033
1034 kfree(hctx->rqs);
1035
1036 if (hctx->tags)
1037 blk_mq_free_tags(hctx->tags);
1038}
1039
1040static size_t order_to_size(unsigned int order)
1041{
1042 size_t ret = PAGE_SIZE;
1043
1044 while (order--)
1045 ret *= 2;
1046
1047 return ret;
1048}
1049
1050static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
1051 unsigned int reserved_tags, int node)
1052{
1053 unsigned int i, j, entries_per_page, max_order = 4;
1054 size_t rq_size, left;
1055
1056 INIT_LIST_HEAD(&hctx->page_list);
1057
1058 hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
1059 GFP_KERNEL, node);
1060 if (!hctx->rqs)
1061 return -ENOMEM;
1062
1063 /*
1064 * rq_size is the size of the request plus driver payload, rounded
1065 * to the cacheline size
1066 */
1067 rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
1068 cache_line_size());
1069 left = rq_size * hctx->queue_depth;
1070
1071 for (i = 0; i < hctx->queue_depth;) {
1072 int this_order = max_order;
1073 struct page *page;
1074 int to_do;
1075 void *p;
1076
1077 while (left < order_to_size(this_order - 1) && this_order)
1078 this_order--;
1079
1080 do {
1081 page = alloc_pages_node(node, GFP_KERNEL, this_order);
1082 if (page)
1083 break;
1084 if (!this_order--)
1085 break;
1086 if (order_to_size(this_order) < rq_size)
1087 break;
1088 } while (1);
1089
1090 if (!page)
1091 break;
1092
1093 page->private = this_order;
Dave Hansen67534712014-01-08 20:17:46 -07001094 list_add_tail(&page->lru, &hctx->page_list);
Jens Axboe320ae512013-10-24 09:20:05 +01001095
1096 p = page_address(page);
1097 entries_per_page = order_to_size(this_order) / rq_size;
1098 to_do = min(entries_per_page, hctx->queue_depth - i);
1099 left -= to_do * rq_size;
1100 for (j = 0; j < to_do; j++) {
1101 hctx->rqs[i] = p;
1102 blk_mq_rq_init(hctx, hctx->rqs[i]);
1103 p += rq_size;
1104 i++;
1105 }
1106 }
1107
1108 if (i < (reserved_tags + BLK_MQ_TAG_MIN))
1109 goto err_rq_map;
1110 else if (i != hctx->queue_depth) {
1111 hctx->queue_depth = i;
1112 pr_warn("%s: queue depth set to %u because of low memory\n",
1113 __func__, i);
1114 }
1115
1116 hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
1117 if (!hctx->tags) {
1118err_rq_map:
1119 blk_mq_free_rq_map(hctx);
1120 return -ENOMEM;
1121 }
1122
1123 return 0;
1124}
1125
1126static int blk_mq_init_hw_queues(struct request_queue *q,
1127 struct blk_mq_reg *reg, void *driver_data)
1128{
1129 struct blk_mq_hw_ctx *hctx;
1130 unsigned int i, j;
1131
1132 /*
1133 * Initialize hardware queues
1134 */
1135 queue_for_each_hw_ctx(q, hctx, i) {
1136 unsigned int num_maps;
1137 int node;
1138
1139 node = hctx->numa_node;
1140 if (node == NUMA_NO_NODE)
1141 node = hctx->numa_node = reg->numa_node;
1142
1143 INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
1144 spin_lock_init(&hctx->lock);
1145 INIT_LIST_HEAD(&hctx->dispatch);
1146 hctx->queue = q;
1147 hctx->queue_num = i;
1148 hctx->flags = reg->flags;
1149 hctx->queue_depth = reg->queue_depth;
1150 hctx->cmd_size = reg->cmd_size;
1151
1152 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1153 blk_mq_hctx_notify, hctx);
1154 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1155
1156 if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
1157 break;
1158
1159 /*
1160 * Allocate space for all possible cpus to avoid allocation in
1161 * runtime
1162 */
1163 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1164 GFP_KERNEL, node);
1165 if (!hctx->ctxs)
1166 break;
1167
1168 num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
1169 hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
1170 GFP_KERNEL, node);
1171 if (!hctx->ctx_map)
1172 break;
1173
1174 hctx->nr_ctx_map = num_maps;
1175 hctx->nr_ctx = 0;
1176
1177 if (reg->ops->init_hctx &&
1178 reg->ops->init_hctx(hctx, driver_data, i))
1179 break;
1180 }
1181
1182 if (i == q->nr_hw_queues)
1183 return 0;
1184
1185 /*
1186 * Init failed
1187 */
1188 queue_for_each_hw_ctx(q, hctx, j) {
1189 if (i == j)
1190 break;
1191
1192 if (reg->ops->exit_hctx)
1193 reg->ops->exit_hctx(hctx, j);
1194
1195 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1196 blk_mq_free_rq_map(hctx);
1197 kfree(hctx->ctxs);
1198 }
1199
1200 return 1;
1201}
1202
1203static void blk_mq_init_cpu_queues(struct request_queue *q,
1204 unsigned int nr_hw_queues)
1205{
1206 unsigned int i;
1207
1208 for_each_possible_cpu(i) {
1209 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1210 struct blk_mq_hw_ctx *hctx;
1211
1212 memset(__ctx, 0, sizeof(*__ctx));
1213 __ctx->cpu = i;
1214 spin_lock_init(&__ctx->lock);
1215 INIT_LIST_HEAD(&__ctx->rq_list);
1216 __ctx->queue = q;
1217
1218 /* If the cpu isn't online, the cpu is mapped to first hctx */
1219 hctx = q->mq_ops->map_queue(q, i);
1220 hctx->nr_ctx++;
1221
1222 if (!cpu_online(i))
1223 continue;
1224
1225 /*
1226 * Set local node, IFF we have more than one hw queue. If
1227 * not, we remain on the home node of the device
1228 */
1229 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1230 hctx->numa_node = cpu_to_node(i);
1231 }
1232}
1233
1234static void blk_mq_map_swqueue(struct request_queue *q)
1235{
1236 unsigned int i;
1237 struct blk_mq_hw_ctx *hctx;
1238 struct blk_mq_ctx *ctx;
1239
1240 queue_for_each_hw_ctx(q, hctx, i) {
1241 hctx->nr_ctx = 0;
1242 }
1243
1244 /*
1245 * Map software to hardware queues
1246 */
1247 queue_for_each_ctx(q, ctx, i) {
1248 /* If the cpu isn't online, the cpu is mapped to first hctx */
1249 hctx = q->mq_ops->map_queue(q, i);
1250 ctx->index_hw = hctx->nr_ctx;
1251 hctx->ctxs[hctx->nr_ctx++] = ctx;
1252 }
1253}
1254
1255struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
1256 void *driver_data)
1257{
1258 struct blk_mq_hw_ctx **hctxs;
1259 struct blk_mq_ctx *ctx;
1260 struct request_queue *q;
1261 int i;
1262
1263 if (!reg->nr_hw_queues ||
1264 !reg->ops->queue_rq || !reg->ops->map_queue ||
1265 !reg->ops->alloc_hctx || !reg->ops->free_hctx)
1266 return ERR_PTR(-EINVAL);
1267
1268 if (!reg->queue_depth)
1269 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1270 else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
1271 pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
1272 reg->queue_depth = BLK_MQ_MAX_DEPTH;
1273 }
1274
Jens Axboe320ae512013-10-24 09:20:05 +01001275 if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
1276 return ERR_PTR(-EINVAL);
1277
1278 ctx = alloc_percpu(struct blk_mq_ctx);
1279 if (!ctx)
1280 return ERR_PTR(-ENOMEM);
1281
1282 hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1283 reg->numa_node);
1284
1285 if (!hctxs)
1286 goto err_percpu;
1287
1288 for (i = 0; i < reg->nr_hw_queues; i++) {
1289 hctxs[i] = reg->ops->alloc_hctx(reg, i);
1290 if (!hctxs[i])
1291 goto err_hctxs;
1292
1293 hctxs[i]->numa_node = NUMA_NO_NODE;
1294 hctxs[i]->queue_num = i;
1295 }
1296
1297 q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
1298 if (!q)
1299 goto err_hctxs;
1300
1301 q->mq_map = blk_mq_make_queue_map(reg);
1302 if (!q->mq_map)
1303 goto err_map;
1304
1305 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1306 blk_queue_rq_timeout(q, 30000);
1307
1308 q->nr_queues = nr_cpu_ids;
1309 q->nr_hw_queues = reg->nr_hw_queues;
1310
1311 q->queue_ctx = ctx;
1312 q->queue_hw_ctx = hctxs;
1313
1314 q->mq_ops = reg->ops;
Jens Axboe94eddfb2013-11-19 09:25:07 -07001315 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
Jens Axboe320ae512013-10-24 09:20:05 +01001316
Christoph Hellwig1be036e2014-02-07 10:22:39 -08001317 q->sg_reserved_size = INT_MAX;
1318
Jens Axboe320ae512013-10-24 09:20:05 +01001319 blk_queue_make_request(q, blk_mq_make_request);
1320 blk_queue_rq_timed_out(q, reg->ops->timeout);
1321 if (reg->timeout)
1322 blk_queue_rq_timeout(q, reg->timeout);
1323
Christoph Hellwig30a91cb2014-02-10 03:24:38 -08001324 if (reg->ops->complete)
1325 blk_queue_softirq_done(q, reg->ops->complete);
1326
Jens Axboe320ae512013-10-24 09:20:05 +01001327 blk_mq_init_flush(q);
1328 blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
1329
Christoph Hellwig18741982014-02-10 09:29:00 -07001330 q->flush_rq = kzalloc(round_up(sizeof(struct request) + reg->cmd_size,
1331 cache_line_size()), GFP_KERNEL);
1332 if (!q->flush_rq)
Jens Axboe320ae512013-10-24 09:20:05 +01001333 goto err_hw;
1334
Christoph Hellwig18741982014-02-10 09:29:00 -07001335 if (blk_mq_init_hw_queues(q, reg, driver_data))
1336 goto err_flush_rq;
1337
Jens Axboe320ae512013-10-24 09:20:05 +01001338 blk_mq_map_swqueue(q);
1339
1340 mutex_lock(&all_q_mutex);
1341 list_add_tail(&q->all_q_node, &all_q_list);
1342 mutex_unlock(&all_q_mutex);
1343
1344 return q;
Christoph Hellwig18741982014-02-10 09:29:00 -07001345
1346err_flush_rq:
1347 kfree(q->flush_rq);
Jens Axboe320ae512013-10-24 09:20:05 +01001348err_hw:
1349 kfree(q->mq_map);
1350err_map:
1351 blk_cleanup_queue(q);
1352err_hctxs:
1353 for (i = 0; i < reg->nr_hw_queues; i++) {
1354 if (!hctxs[i])
1355 break;
1356 reg->ops->free_hctx(hctxs[i], i);
1357 }
1358 kfree(hctxs);
1359err_percpu:
1360 free_percpu(ctx);
1361 return ERR_PTR(-ENOMEM);
1362}
1363EXPORT_SYMBOL(blk_mq_init_queue);
1364
1365void blk_mq_free_queue(struct request_queue *q)
1366{
1367 struct blk_mq_hw_ctx *hctx;
1368 int i;
1369
1370 queue_for_each_hw_ctx(q, hctx, i) {
Jens Axboe320ae512013-10-24 09:20:05 +01001371 kfree(hctx->ctx_map);
1372 kfree(hctx->ctxs);
1373 blk_mq_free_rq_map(hctx);
1374 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1375 if (q->mq_ops->exit_hctx)
1376 q->mq_ops->exit_hctx(hctx, i);
1377 q->mq_ops->free_hctx(hctx, i);
1378 }
1379
1380 free_percpu(q->queue_ctx);
1381 kfree(q->queue_hw_ctx);
1382 kfree(q->mq_map);
1383
1384 q->queue_ctx = NULL;
1385 q->queue_hw_ctx = NULL;
1386 q->mq_map = NULL;
1387
1388 mutex_lock(&all_q_mutex);
1389 list_del_init(&q->all_q_node);
1390 mutex_unlock(&all_q_mutex);
1391}
Jens Axboe320ae512013-10-24 09:20:05 +01001392
1393/* Basically redo blk_mq_init_queue with queue frozen */
Paul Gortmakerf618ef72013-11-14 08:26:02 -07001394static void blk_mq_queue_reinit(struct request_queue *q)
Jens Axboe320ae512013-10-24 09:20:05 +01001395{
1396 blk_mq_freeze_queue(q);
1397
1398 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1399
1400 /*
1401 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1402 * we should change hctx numa_node according to new topology (this
1403 * involves free and re-allocate memory, worthy doing?)
1404 */
1405
1406 blk_mq_map_swqueue(q);
1407
1408 blk_mq_unfreeze_queue(q);
1409}
1410
Paul Gortmakerf618ef72013-11-14 08:26:02 -07001411static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1412 unsigned long action, void *hcpu)
Jens Axboe320ae512013-10-24 09:20:05 +01001413{
1414 struct request_queue *q;
1415
1416 /*
1417 * Before new mapping is established, hotadded cpu might already start
1418 * handling requests. This doesn't break anything as we map offline
1419 * CPUs to first hardware queue. We will re-init queue below to get
1420 * optimal settings.
1421 */
1422 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1423 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1424 return NOTIFY_OK;
1425
1426 mutex_lock(&all_q_mutex);
1427 list_for_each_entry(q, &all_q_list, all_q_node)
1428 blk_mq_queue_reinit(q);
1429 mutex_unlock(&all_q_mutex);
1430 return NOTIFY_OK;
1431}
1432
1433static int __init blk_mq_init(void)
1434{
Jens Axboe320ae512013-10-24 09:20:05 +01001435 blk_mq_cpu_init();
1436
1437 /* Must be called after percpu_counter_hotcpu_callback() */
1438 hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
1439
1440 return 0;
1441}
1442subsys_initcall(blk_mq_init);